We have described a new recommendation technique based on several measurements. In
addition to traditional measurements support and confidence, we also test the effectiveness of a novel measurement - information value - which balances the gain between current and future
selections. Its performance surpasses that of confidence, and it is still computable in real time.
Such a system is a complement to expert systems and traditional practice guidelines, and can be
very useful for nursing education and clinical quality control.
To date we have only experimented with relationships among diagnoses. In future work we will
also examine the relationships among diagnoses, outcomes, and interventions. The effectiveness
difference between expert systems and such a recommender is also interesting. Rules from
experts’ knowledge could be more accurate but they are not easily updated and specialized for
different hospitals. Can we combine these two kinds of systems to achieve better results?
Another promising direction is to incorporate contextual information into the recommendation
process and make recommendations based on multiple dimensions, patient profiles, and other
information [1]. Finally, we will examine the trade-off between confidence (immediate
probability) and entropy (future probability) in the information value measurement, and adjust it
to perform better on specific problems.